前面兩篇帖子分別總結了innodb_thread_concurrency和thread pool的原理：

前者是在儲存引擎層面限制併發執行的執行緒數，程式碼路徑過於靠後，此時query已在server層完成解析；

後者則是在server層建立多組常駐執行緒，用於接收客戶端連線傳送的query並代為執行，而不是為每個連線單獨建立一個執行緒。

除了這兩種解決方案，還可以在server層進行running thread數量判斷，如果達到閾值則直接報錯或sleep。

下面介紹一下其實現原理和patch原始碼，資料來源。

thread_running的意義

thread_running狀態變數記錄了當前併發執行stmt/command的數量，執行前加1執行後減1；

程式碼邏輯

do_command

--&gtdispatch_command

...

inc_thread_running

...

mysql_execute_command or execute_some_command

...

dec_thread_running

...

Thread_running突然飆高的誘因：

1 客戶端連線暴增；

2 系統效能瓶頸，如CPU,IO或者mem swap；

3 異常sql；

往往在這種情況下，MySQL server會表現出hang住的假象。

解決方案

暫時禁止新sql執行，為此引入兩個閾值low_watermark和high_watermark，以及變數threads_running_ctl_mode(selects或者all )；

執行query前，檢查thread_running，

1 若其已達high_watermark閾值則直接拒絕執行並返回錯誤：mysql server is too busy

2 若其位於low和high之間，則sleep 5ms，然後繼續嘗試，累計等待100ms後則執行

3 對於已經開啟事務和super使用者，不做限制

4 threads_running_ctl_mode控制query型別：SELECTS/ALL，預設為SELECTS，表示隻影響SELECT語句

Patch部分原始碼見注1

進一步改進

將低水位限流從sleep-retry最佳化為基於FIFO的cond-wait/signal(實現8個FIFO)；

1 高水位限流(這點保持不變)；

2 低水位最佳化；其他解決方案：mariadb開發thread pool，percona在其上實現了優先佇列；

本patch優勢：思路與thread pool一致，但程式碼更簡潔(不到1000行)；而且增加了特定query的過濾；

Patch部分程式碼見注2

低水位最佳化細節

1 新增thread_active記錄併發執行緒數，位於mysql_execute_command(sql解析之後)，高水位則在query解析之前判斷；

Thread_active只統計select/DML，而commit/rollback則放過。

2 採用FIFO，當thread_active >= thread_running_low_watermark時程式進入FIFO等待，其他執行緒執行完sql後喚醒FIFO；

保證併發執行緒控制在thread_running_low_watermark內，同時引入threads_running_wait_timeout控制執行緒在FIFO最大等待時間，超時則直接報錯返回。

3 引入8個FIFO，降低了進出FIFO的鎖競爭，執行緒採用RR分配到不同fifo，每個佇列限制併發執行執行緒為threads_running_low_watermark/8。

已經透過高水位驗證的thread，開始執行query，[解析後進行低水位判斷，若透過則執行]，執行當前sql完畢後，thread可能發起新query，則重複[]過程。

新增系統變數

threads_running_wait_timeout：進入FIFO排隊最長時間，等待超時後sql被拒，預設100，單位為毫秒ms。

新增狀態變數

threads_active: 當前併發SELECT/INSERT/UPDATE/DELETE執行的執行緒數目；

threads_wait：當前進入到FIFO中等待的執行緒數目；

測試效果

./sysbench --test=tests/db/select.lua --max-requests=0 --mysql-host=myxxxx.cm3 --mysql-user=test --mysql-table-engine=innodb --oltp-table-size=5000000 --oltp-tables-count=32

normal mysql-0 : 未打補丁版本，設定innodb_thread_concurrency=0

normal mysql-1 : 未打補丁版本，innodb_thread_concurrency=32

patched mysql : 低水位限流補丁版本（活躍執行緒數不超過64）

注1

+static my_bool thread_running_control(THD *thd, ulong tr)
+{
+ int slept_cnt= 0;
+ ulong tr_low, tr_high;
+ DBUG_ENTER("thread_running_control");
+
+ /*
+ Super user/slave thread will not be affected at any time,
+ transactions that have already started will continue.
+ */
+ if ( thd->security_ctx->master_access & SUPER_ACL|| --對於super許可權的使用者和已經開啟的事務不做限制
+ thd->in_active_multi_stmt_transaction() ||
+ thd->slave_thread)
+ DBUG_RETURN(FALSE);
+
+ /*
+ To promise that tr_low will never be greater than tr_high,
+ as values may be changed between these two statements.
+ eg.
+ (low, high) = (200, 500)
+ 1. read low = 200
+ 2. other sessions: set low = 20; set high = 80
+ 3. read high = 80
+ Don't take a lock here to avoid lock contention.
+ */
+ do
+ {
+ tr_low= thread_running_low_watermark;
+ tr_high= thread_running_high_watermark;
+
+ } while (tr_low > tr_high);
+
+check_buzy:
+
+ /* tr_high is promised to be non-zero.*/
+ if ((tr_low == 0 && tr < tr_high) || (tr_low != 0 && tr < tr_low))
+ DBUG_RETURN(FALSE);
+
+ if (tr >= tr_high)
+ {
+ int can_reject= 1;
+
+ /* thread_running_ctl_mode: 0 -> SELECTS, 1 -> ALL. */
+ if (thread_running_ctl_mode == 0)
+ {
+ int query_is_select= 0;
+ if (thd->query_length() >= 8)
+ {
+ char *p= thd->query(); --讀取query text的前6個字元，以判斷是否為select
+ if (my_toupper(system_charset_info, p[0]) == 'S' &&
+ my_toupper(system_charset_info, p[1]) == 'E' &&
+ my_toupper(system_charset_info, p[2]) == 'L' &&
+ my_toupper(system_charset_info, p[3]) == 'E' &&
+ my_toupper(system_charset_info, p[4]) == 'C' &&
+ my_toupper(system_charset_info, p[5]) == 'T')
+
+ query_is_select= 1;
+ }
+
+ if (!query_is_select)
+ can_reject= 0;
+ }
+
+ if (can_reject)
+ {
+ inc_thread_rejected();
+ DBUG_RETURN(TRUE);
+ }
+ else
+ DBUG_RETURN(FALSE);
+ }
+
+ if (tr_low != 0 && tr >= tr_low)
+ {
+ /*
+ If total slept time exceed 100ms and thread running does not
+ reach high watermark, let it in.
+ */
+ if (slept_cnt >= 20)
+ DBUG_RETURN(FALSE);
+
+ dec_thread_running()
+
+ /* wait for 5ms. */
+ my_sleep(5000UL);
+
+ slept_cnt++;
+ tr= inc_thread_running() - 1;
+
+ goto check_buzy;
+ }
+
+ DBUG_RETURN(FALSE);
+}
+
+/**
Perform one connection-level (COM_XXXX) command.
@param command type of command to perform
@@ -1016,7 +1126,8 @@
thd->set_query_id(get_query_id());
if (!(server_command_flags[command] & CF_SKIP_QUERY_ID))
next_query_id();
- inc_thread_running();
+ /* remember old value of thread_running for *thread_running_control*. */
+ int32 tr= inc_thread_running() - 1;
if (!(server_command_flags[command] & CF_SKIP_QUESTIONS))
statistic_increment(thd->status_var.questions, &LOCK_status);

@@ -1129,6 +1240,13 @@
{
if (alloc_query(thd, packet, packet_length))
break; // fatal error is set
+
+ if (thread_running_control(thd, (ulong)tr))
+ {
+ my_error(ER_SERVER_THREAD_RUNNING_TOO_HIGH, MYF(0));
+ break;
+ }
+
MYSQL_QUERY_START(thd->query(), thd->thread_id, (char *) (thd->db ? thd->db : ""), &thd->security_ctx->priv_user[0])

注2

+/**
Perform one connection-level (COM_XXXX) command.

@param command type of command to perform
@@ -1177,7 +1401,7 @@
command= COM_SHUTDOWN;
}
thd->set_query_id(next_query_id());
- inc_thread_running();
+ int32 tr= inc_thread_running();

if (!(server_command_flags[command] & CF_SKIP_QUESTIONS))
statistic_increment(thd->status_var.questions, &LOCK_status);
@@ -1209,6 +1433,15 @@
goto done;
}

+ if (command == COM_QUERY && alloc_query(thd, packet, packet_length))
+ goto endof_case; // fatal error is set
+
+ if (thread_running_control_high(thd, tr))
+ {
+ my_error(ER_SERVER_THREAD_RUNNING_TOO_HIGH, MYF(0));
+ goto endof_case;
+ }
+
switch (command) {
case COM_INIT_DB:
{
@@ -1311,8 +1544,6 @@
}
case COM_QUERY:
{
- if (alloc_query(thd, packet, packet_length))
- break; // fatal error is set
MYSQL_QUERY_START(thd->query(), thd->thread_id,
(char *) (thd->db ? thd->db : ""),
&thd->security_ctx->priv_user[0],
@@ -1751,6 +1982,7 @@
my_message(ER_UNKNOWN_COM_ERROR, ER(ER_UNKNOWN_COM_ERROR), MYF(0));
break;
}
+endof_case:

done:
DBUG_ASSERT(thd->derived_tables == NULL &&
@@ -2502,12 +2734,37 @@
Opt_trace_array trace_command_steps(&thd->opt_trace, "steps");

DBUG_ASSERT(thd->transaction.stmt.cannot_safely_rollback() == FALSE);
+ bool count_active= false;

if (need_traffic_control(thd, lex->sql_command))
{
thd->killed = THD::KILL_QUERY;
goto error;
}
+
+ switch (lex->sql_command) {
+
+ case SQLCOM_SELECT:
+ case SQLCOM_UPDATE:
+ case SQLCOM_UPDATE_MULTI:
+ case SQLCOM_DELETE:
+ case SQLCOM_DELETE_MULTI:
+ case SQLCOM_INSERT:
+ case SQLCOM_INSERT_SELECT:
+ case SQLCOM_REPLACE:
+ case SQLCOM_REPLACE_SELECT:
+ count_active= true;
+ break;
+ default:
+ break;
+ }
+
+ if (count_active && thread_running_control_low_enter(thd))
+ {
+ my_error(ER_SERVER_THREAD_RUNNING_TOO_HIGH, myf(0));
+ goto error;
+ }
+
status_var_increment(thd->status_var.com_stat[lex->sql_command]);

switch (gtid_pre_statement_checks(thd))
@@ -4990,6 +5247,9 @@

finish:

+ if (count_active)
+ thread_running_control_low_exit(thd);
+
DBUG_ASSERT(!thd->in_active_multi_stmt_transaction() ||
thd->in_multi_stmt_transaction_mode());

+static my_bool thread_running_control_high(THD *thd, int32 tr)
+{
+ int32 tr_high;
+ DBUG_ENTER("thread_running_control_high");
+
+ tr_high= (int32)thread_running_high_watermark;
+
+ /* thread_running_ctl_mode: 0 -> SELECTS, 1 -> ALL. */
+ if ((!tr_high || tr <= tr_high) ||
+ thd->transaction.is_active() ||
+ thd->get_command() != COM_QUERY ||
+ thd->security_ctx->master_access & SUPER_ACL ||
+ thd->slave_thread)
+ DBUG_RETURN(FALSE);
+
+ const char *query= thd->query();
+ uint32 len= thd->query_length();
+
+ if ((!has_prefix(query, len, "SELECT", 6) && thread_running_ctl_mode == 0) || --不再是逐個字元判斷
+ has_prefix(query, len, "COMMIT", 6) ||
+ has_prefix(query, len, "ROLLBACK", 8))
+ DBUG_RETURN(FALSE);
+
+ /* confirm again*/
+ if (tr > tr_high && get_thread_running() > tr_high)
+ {
+ __sync_add_and_fetch(&thread_rejected, 1);
+ DBUG_RETURN(TRUE);
+ }
+
+ DBUG_RETURN(FALSE);
+}
+

+static my_bool thread_running_control_low_enter(THD *thd)
+{
+ int res= 0;
+ int32 tr_low;
+ my_bool ret= FALSE;
+ my_bool slept= FALSE;
+ struct timespec timeout;
+ Thread_conc_queue *queue;
+ DBUG_ENTER("thread_running_control_low_enter");
+
+ /* update global status */
+ __sync_add_and_fetch(&thread_active, 1);
+
+ tr_low= (int32)queue_tr_low_watermark;
+ queue= thread_conc_queues + thd->query_id % N_THREAD_CONC_QUEUE;
+
+ queue->lock();--問1：在進行低水位判斷前，先鎖定FIFO，避免低水位驗證失敗時無法獲取FIFO鎖進而不能放入FIFO；
+
+retry:
+
+ if ((!tr_low || queue->thread_active < tr_low) ||
+ (thd->lex->sql_command != SQLCOM_SELECT && thread_running_ctl_mode == 0) ||
+ (!slept && (thd->transaction.is_active() ||
+ thd->security_ctx->master_access & SUPER_ACL || thd->slave_thread)))
+ {
+ queue->thread_active++; --判斷是否滿足進入FIFO條件，如不滿足則立即更新thread_active++，解鎖queue並退出；
+ queue->unlock();
+ DBUG_RETURN(ret);
+ }
+
+ if (!slept)
+ {
+ queue->unlock();
+
+ /* sleep for 500 us */
+ my_sleep(500);
+ slept= TRUE;
+ queue->lock();
+
+ goto retry;
+ }
+
+ /* get a free wait-slot */
+ Thread_wait_slot *slot= queue->pop_free();
+
+ /* can't find a free wait slot, must let the query enter */
+ if (!slot)-- 當FIFO都滿了，即無法把當前執行緒放入，則必須放行讓該sql正常執行
+ {
+ queue->thread_active++;
+ queue->unlock();
+ DBUG_RETURN(ret);
+ }
+
+ slot->signaled= false;
+ slot->wait_ended= false;
+
+ /* put slot into waiting queue. */
+ queue->push_back_wait(slot);
+ queue->thread_wait++;
+
+ queue->unlock();
+
+ /* update global status */
+ thd_proc_info(thd, "waiting in server fifo");
+ __sync_sub_and_fetch(&thread_active, 1);
+ __sync_add_and_fetch(&thread_wait, 1);
+
+ /* cond-wait for at most thread_running_wait_timeout(ms). */
+ set_timespec_nsec(timeout, thread_running_wait_timeout_ns);
+
+ mysql_mutex_lock(&slot->mutex);
+ while (!slot->signaled)
+ {
+ res= mysql_cond_timedwait(&slot->cond, &slot->mutex, &timeout);
+ /* no need to signal if cond-wait timedout */
+ slot->signaled= true;
+ }
+ mysql_mutex_unlock(&slot->mutex);
+
+ queue->lock();
+ queue->thread_wait--;
+ queue->thread_active++;
+
+ /* remove slot from waiting queue. */
+ queue->remove_wait(slot);
+ /* put slot into the free queue for reuse. */
+ queue->push_back_free(slot);
+
+ queue->unlock();
+
+ /* update global status */
+ __sync_sub_and_fetch(&thread_wait, 1);
+ __sync_add_and_fetch(&thread_active, 1);
+ thd_proc_info(thd, 0);
+
+ if (res == ETIMEDOUT || res == ETIME)
+ {
+ ret= TRUE; // indicate that query is rejected.
+ __sync_add_and_fetch(&thread_rejected, 1);
+ }
+
+ DBUG_RETURN(ret);
+}

mysql併發執行緒控制之控制thread_running數量

相關文章